Comparison of choice models representing correlation and random taste variation: An application to airline passengers' rescheduling behavior

There has been increasing interest in using mixed multinomial logit models because they enable random taste variation by allowing parameters of the utility function to vary across individuals and have been shown to theoretically approximate any random utility model via the inclusion of an appropriate set of error components. However, this study provides a motivation for integrating complex closed-form models to represent correlation among alternatives with random coefficients. Using the nested logit (NL) model as an example, we demonstrate that while a mixed model can approximate a NL model theoretically, empirically the NL model may not be well approximated by its mixture analog. Specifically, it is difficult to obtain precise correlation estimates from mixture models, particularly for nests that have small correlations.; A second theoretical contribution of this work relates to efficient estimation of nested logit models for choice-based samples. Currently, the benefit of collecting choice-based samples diminishes when modeling consumers' behavior using NL models because of the need to use consistent estimators that are often inefficient and/or complicated to implement. In contrast, benefits of using choice-based samples are retained when using the simple multinomial logit model because, under conditions that are relatively easy to satisfy in practice, the exogenous sample maximum likelihood estimator can be used. This study shows the exogenous sample maximum likelihood estimator can also be used with choice-based samples for NL models.; Finally, this work contributes to the state-of-the-art practice in airline yield management. This work represents the most comprehensive study of airline passengers' day of departure rescheduling behavior and is the first to examine standby behavior. The model explores differences in show, no-show, and standby behavior due to booking class, frequent flier status, schedule attributes of the carrier of interest and its competitors differentiated by level of service, and other factors. Particular emphasis is placed on understanding how standby behavior differs between outbound versus inbound itineraries and how behavior changed after the terrorist attacks of September 11, 2001. A validation analysis demonstrates benefits of incorporating passenger data and outbound/inbound itinerary information in airline no-show models.